https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101024
Bug ID: 101024
Summary: Missed min expression at phiopt1
Product: gcc
Version: 12.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: enhancement
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: pinskia at gcc dot gnu.org
Target Milestone: ---
Take:
unsigned f(int nunits)
{
int units_per = 32 / (8);
unsigned int chunk_nunits = ((nunits) < (units_per) ? (nunits) :
(units_per));
return chunk_nunits;
}
-------- CUT ------
Compile with the C++ front-end and -O2 (note C++ is needed as fold won't
convert ?: into min/max expression due to lvalue requirements).
You will notice this is not caught by phiopt1 even though it should be while it
is caught now (on the trunk) by phiopt2 (match-and-simplify). That is the
minmax_replacement does not handle the above case but match does.
Note this shows up in read code (GCC) in assemble_real (varasm.c). There are
two cases of this in assemble_real but only one of the two is caught in phiopt2
though because there looks like some jump threading that had happened before
that causes other issues.
I am filing this as a bug only so I can keep track of moving minmax_replacement
to match-and-simplify.
